Depth Image Inpainting via Single Depth Features Learning
Junbo Mao, Jupeng Li, Feng Li, Chengkai Wan · 2020
The holes in depth images, which are one of the main factors affecting image quality and applications. In order to fill holes effectively for the depth images, this paper proposes a single depth image inpainting method based on deep learning. First, features are learned and extracted from a single depth image by U-Net network. Then, the similarity measurement loss between the output and ground truth is back-propagated to update network parameters. Finally, the depth image inpainting result is created by combing the output hole filling result and original image. The NYUv2 dataset was used for the training and evaluation. Experimental results show that results of our method which only learns feature information from single depth image are similar to the optimal results of various hole filling methods which using color image information.